Michigan
Retirement
Research
University of
Working Paper
WP 2009-207
Center
Work Disability, Work, and Justification Bias in Europe and the
U.S.
Arie Kapteyn, James P. Smith and Arthur van Soest
MR
RC
Project #: UM09-15
Work Disability, Work, and Justification Bias in Europe and the
U.S.
Arie Kapteyn
RAND
James P. Smith
RAND
Arthur van Soest
Tilburg University, Netspar & RAND
August 2009
Michigan Retirement Research Center
University of Michigan
P.O. Box 1248
Ann Arbor, MI 48104
http://www.mrrc.isr.umich.edu/
(734) 615-0422
Acknowledgements
This work was supported by a grant from the Social Security Administration through the
Michigan Retirement Research Center (Grant # 10-M-98362-5-01). The findings and
conclusions expressed are solely those of the author and do not represent the views of the Social
Security Administration, any agency of the Federal government, or the Michigan Retirement
Research Center.
Regents of the University of Michigan
Julia Donovan Darrow, Ann Arbor; Laurence B. Deitch, Bingham Farms; Denise Ilitch, Birmingham; Olivia P.
Maynard, Goodrich; Andrea Fischer Newman, Ann Arbor; Andrew C. Richner, Grosse Pointe Park; S. Martin
Taylor, Gross Pointe Farms; Katherine E. White, Ann Arbor; Mary Sue Coleman, ex officio
Work Disability, Work, and Justification Bias in Europe and the
U.S.
Abstract
To analyze the effect of health on work, many studies use a simple self-assessed health measure
based upon a question such as “do you have an impairment or health problem limiting the kind
or amount of work you can do?” A possible drawback of such a measure is the possibility that
different groups of respondents may use different response scales. This is commonly referred to
as “differential item functioning” (DIF). A specific form of DIF is justification bias: to justify the
fact that they don’t work, non-working respondents may classify a given health problem as a
more serious work limitation than working respondents. In this paper we use anchoring vignettes
to identify justification bias and other forms of DIF across countries and socio-economic groups
among older workers in the U.S. and Europe. Generally, we find differences in response scales
across countries, partly related to social insurance generosity and employment protection.
Furthermore, we find significant evidence of justification bias in the U.S. but not in Europe,
suggesting differences in social norms concerning work.
Authors’ Acknowledgements
JEL codes: J28, I12, C81
Keywords: Work limiting disability, Vignettes, Reporting bias
This research was supported by grants from the National Institute on Aging to RAND and the
Social Security Administration through MRRC. It uses data from SHARE Wave 1, primarily
funded by the European Commission through its 5th and 6th framework programmes (project
numbers QLK6-CT-2001-00360; RII-CT- 2006-062193; CIT5-CT-2005-028857) and of HRS,
funded by NIA. We are grateful to Angus Deaton for useful comments on an earlier version.
1. Introduction
The fraction of workers on disability insurance (DI) is vastly different across Western
European countries with similar levels of economic development and comparable access to
modern medical technology and treatment (Eurostat, 2001). Institutional differences in eligibility
rules or generosity of benefits contribute to an explanation of differences in disability rolls
(Boersch-Supan, 2007). Recent survey data show, however, that significant differences between
Western European countries and the U.S. are also found in self-reports of work limiting
disabilities. Table 1 below, taken from Kapteyn et al. (2009), illustrates the point.1 We see
considerable variation in both DI-expenditures as a percentage of GDP and the percentage of
males between 40 and 65 reporting some form of work disability. Remarkably the two columns
are only weakly correlated: the correlation is 0.20.
Table 1: Expenditures on Disability Insurance and
Self-reported Male Work Disability, 2001
Germany
Denmark
Netherlands
Belgium
France
UK
Ireland
Italy
Greece
Spain
Portugal
Austria
Finland
U.S.
DI expenditure as a %
of GDP
Self-reported male work disability,
40-65, 2001 (%)
1.6
2.7
4
2.2
1.7
2.2
1.3
2
1.6
2.3
2.4
2.3
3.1
1.1
40.3
22.0
24.5
14.3
20.5
13.1
15.7
8.0
13.3
15.5
22.9
17.8
29.0
19.3
1
Self-reports for the European countries are taken from the European Community Household Panel. For the U.S.
they are taken from the Panel Study of Income Dynamics. The exact question on work disability in ECHP is: “Are
you hampered in your daily activities by any physical or mental health problem, illness or disability?” In the PSID, it
is: “Do you have any physical or nervous condition that limits the type of work or the amount of work you can do?”
2
Croda and Skinner (2009), using data from SHARE and HRS, find little or no evidence
that in countries with a higher proportion of individuals on DI, the fraction of DI recipients with
self-reported fair or poor health is lower. This is what one would expect if access to DI were
largely driven by health considerations and if the distribution of health in different countries
would be roughly the same. In that case a system with strict eligibility rules or less generous
benefits would mainly select individuals with the worst health condition. The more generous
benefits or eligibility rules become, the more likely it is that individuals in better health are
drawn into the pool of DI recipients. One explanation for the weak relation between DIrecipiency rates and self-reported health across countries may lie in differences in response
scales used to answer subjective health questions adopted by residents of different countries.
This paper applies the same vignette methodology as in Kapteyn, Smith, and Van Soest
(2007) to determine the extent to which differences in self-reported work-limiting disability
between several European countries and the U.S. are due to differences in response scales. In
addition, we also consider the effect of work disability on employment and the potential effect of
justification bias on the estimate of this effect.
The remainder of the paper is organized as follows. The second section describes the data
collection efforts in HRS and SHARE that make this research possible. Section 3 outlines the
vignette methodology and our statistical model that corrects for response scale differences across
countries. The fourth section briefly describes the vignettes used in this study and how
respondents in different European countries and the US respond to the same vignette scenarios.
Section 5 presents the empirical results and their implications for interpreting observed
differences in self-reported work disability. Section 6 extends the model developed in Section 3
to include employment and justification bias. Section 7 presents the estimation results of that
3
model and shows their economic importance by means of a number of simulations. Our principal
conclusions are contained in the last section.
2. Data Sources
For the United States, we use the 2004 wave of the Health and Retirement Study (HRS),
a bi-annual panel with a representative sample of the US population over 50. It has been
conducted by the University of Michigan since1992. Information collected includes physical and
mental health, socio-economic status (including measures of labor market status, income,
education and wealth), social support, etc. The surveys use a mixture of modes with most new
interviews conducted face-to-face and most re-interviews by phone. Wave-specific overall
response rates for the HRS have improved from 81.7% in 1992 to close to 90% at later waves,
specifically 87.8% in 2004. The survey has a complex sample design and oversamples Blacks,
Hispanics, and residents of Florida. Details on the HRS methodology and the 2004 Wave are
available elsewhere (Heeringa and Connor, 1995). For our analysis we use a subsample of
respondents who first completed a face-to-face interview and later completed a leave-behind
questionnaire. The leave-behind questionnaire consisted of a series of work disability vignettes
and was targeted toward respondents less than 75 years of age.
SHARE is a large-scale project that aims to collect interdisciplinary longitudinal data on
European citizens of age 50 and older and their spouses. The eleven participating countries in the
baseline wave were Denmark, Sweden, Germany, Belgium, the Netherlands, Austria,
Switzerland, Spain, Italy, Greece and France.2 Using a common instrument, SHARE includes
information on physical and mental health, socio-economic status including measures of income,
2
The Czech Republic, Poland, Israel, and Ireland, were added later. See www.share-project.org for exact sample
sizes.
4
education and wealth, and social support. The first wave of the main survey was fielded in 2004,
with between 1000 and 4000 individuals in each country, adding up to about 29,000 individuals.
While containing many unique features, in order to facilitate additional comparisons with
the USA and England, SHARE was purposely modeled after the HRS and the English
Longitudinal Survey of Aging (ELSA) and follows a common set-up across all countries with the
goal of facilitating cross-country research.
For a subset of countries that agreed to participate, SHARE included a set of selfassessments and vignette questions on general health status and on work limiting disabilities as
part of a drop-off questionnaire. The eight countries that participated in this vignette experiment
were Germany, France, Spain, Belgium, Greece, Italy, the Netherlands, and Sweden. The work
disability vignettes were identical to the work disability vignettes in the HRS leave-behind
questionnaire. Both were taken from the surveys used by Kapteyn et al. (2007).
The work disability vignettes deal with work-limiting health problems in the domains of
pain, depression, and cardio-vascular disease. An example of a vignette is the following one (the
first vignette cited in Appendix B which contains the full set of vignettes):
[Eva] feels worried all the time. She gets depressed once a week at work for a couple of days in a
row, thinking about what could go wrong and that her boss will disapprove of her condition. But
she is able to come out of this mood if she concentrates on something else.
For each vignette, the respondent is asked: “Please give us your judgment on how limited these
people are in the kind or amount of work they can do”3
1.
2.
3.
4.
5.
Not at all limited
Mildly limited
Moderately limited
Severely limited
Cannot do any work
3
This is the question wording in the HRS. The wording in SHARE is slightly different (and translated in each
country’s language). See www.share-project.org.
5
The names used in the vignettes (Eva in the example) vary by vignette and, moreover, to each
vignette either a male or a female name was randomly assigned. (See Appendix B for details4)
Preceding the vignette questions, a respondent is asked the self-assessment of work
disability question: “To what extent are you limited in the kind or amount of work you can do
because of an impairment or health problem?”5 with the same answer categories as above. As
explained in the next section, these vignettes make it possible to analyze cross-country
differences in work related health, corrected for international differences in response scales.
Moreover, the vignettes will also be used for comparisons of different socio-economic groups.
3. The Theory of Vignettes
In this section, we first provide an intuitive description of the use of vignettes for identifying
response scale differences and then sketch our statistical approach. The basic idea is illustrated in
Figure 1, which presents the distribution of health related work limitations (work disability for
short) in two hypothetical countries. The density of the continuous work disability variable in
country A is to the left of that in country B, implying that on average, people in country A suffer
less work disability than in country B. The people in the two countries, however, use very
different response scales if asked to report their work limitations on a five-point scale (nonemild-moderate-severe-extreme). These differences may be caused by cultural differences, or
simply be the result of inadequate translation, for instance because there exist no exact one-toone translations of concepts from one language to the other. In the example in the figure, people
in country A have a much more negative view on their capacity for work than people in country
B. Someone in country A with the health indicated by the dashed line would report to have a
4
There we also describe some differences in the formulation of the self-reported disability questions and one
vignette across the two surveys. In this version of the paper we have ignored these differences.
5
Again, this is the wording in the HRS; the wording in SHARE is somewhat different: “Do you have any
impairment or health problem that limits the kind or amount of work you can do?” with answer categories “None”,
“Mild”, “Moderate”, “Severe” and “Extreme.”
6
severe work disability, while a person in country B with the same actual work limitation would
report only a mild work disability The frequency distribution of the self-reports in the two
countries would suggest that people in country A are more work disabled than those in country
B—the opposite of the true disability distribution. Correcting for the differences in the response
scales (DIF, “differential item functioning,” in the terminology of King et al., 2004) is essential
to compare the distributions of actual work limitations in the two countries.
Figure 1. Comparing self-reported health related work limitations
Country A
None
Mild Mod- Severe
erate
Extreme
Work Disability
Country B
None
Mild
Moderate
Severe
Extreme
Work Disability
Vignettes can be used to do the correction. A vignette question describes the work
limitations of a hypothetical person and then asks the respondent to evaluate the work disability
of that person on the same five-point scale that was used for the self-report of their own health.
Since the vignette descriptions are the same in the two countries, the vignette persons in the two
7
countries have the same actual work limitations. For example, respondents can be asked to
evaluate the work limitation of a person whose disability is given by the dashed line. In country
A, this will be evaluated as “severe.” In country B, the evaluation would be “mild.” Since the
actual work disability is the same in the two countries, the difference in the country evaluations
must be due to DIF.
Vignette evaluations thus help to identify differences between the response scales. Using the
scales in one of the two countries as the benchmark, the distribution of evaluations in the other
country can be adjusted by evaluating them on the benchmark scale. The corrected distribution
of the evaluations can then be compared to that in the benchmark country—they are now on the
same scale. In the example in the figure, this will lead to the correct conclusion that people in
country B are more disabled than those in country A, on average. The main identifying
assumption is response consistency: a given respondent uses the same scale for self-reports and
vignette evaluations. King et al. (2004) provide evidence supporting this assumption by
comparing self-reports and vignette evaluations of vision with an objective measure of vision.
Van Soest et al. (2007) provide similar supporting evidence by comparing self-reported drinking
problems to actual alcohol consumption.
We will apply the vignette approach to work limiting disability, using vignettes not only to
obtain international comparisons corrected for DIF, but also for comparisons of different groups
within a given country. For example, it is often hypothesized that men self report themselves in
better health than objective circumstances would warrant, that as they age people adjust their
norms downward about what constitutes good health, and that some of the SES health gradient
reflects different health thresholds by SES rather than true health differences. Vignettes offer the
potential for systematic testing of these hypotheses.
8
3.1 A Formal Model of Response Scales and Vignette Corrections
Our model is an extension of the conditional hopit model (Chopit, cf. King et al., 2004, and
Kapteyn et al., 2007). It explains respondents’ self-reports on work limitations and their reports
on work limitations of hypothetical vignette persons. The first of these is the answer Yi (where i
indicates respondent i) to the question
“Do you have any impairment or health problem that limits the kind or amount of paid work
you can do?”
The questions on work limitations of the vignette persons use the same 5-point scale and
are formulated in the same way (“Does Mr/Mrs X have any impairment or health problem that
limits the type or amount of paid work that he/she can do?”). The answers will be denoted by Yli
where each respondent i evaluates L vignettes l=1,…,L.
Self-reports are modeled as a function of respondent characteristics Xi and an error term
ε i by the following ordered response equation:
Yi* = X i β + ε i ; ε i ∼ N (0,1), ε i independent of X i
(1)
Yi = j if τ i j −1 < Yi * ≤ τ i j , j = 1,...5
(2)
The thresholds τ ij between the categories are given by
τ i0 = −∞, τ i5 = ∞, τ i1 = X iγ 1 + ui , τ i j = τ i j −1 + exp( X iγ j ), j = 2, 3, 4
(3)
ui ∼ N (0, σ u2 ), independent of ε i and X i
(4)
The error term ui reflects unobserved heterogeneity in the thresholds. The fact that different
respondents can use different response scales τ i j is called “differential item functioning” (DIF).
9
Using the self-reports on own work disabilities only, the parameters β and γ 1 cannot be
separately identified;6 the reported outcome only depends on these parameters through their
difference. For example, consider country dummies: two people (with the same characteristics)
in two different countries can have systematically different work disability, but if the scales on
which they report their work disability can also differ across countries, then self-reports alone are
not enough to identify the work disability difference between the countries. This was illustrated
in Figure 1 above.
For our analysis we are using a common set of L=9 vignette questions, three in each of
three domains: affect, pain, and heart problems.7 The evaluations of vignettes l=1,…, L are
modeled using similar ordered response equations:
Yli* = θl + ε li
(5)
Yli = j if τ i j −1 < Yli* ≤ τ i j , j = 1,...5
(6)
ε li ∼ N (0, σ v2 ), independent of each other, of ε i and of X i
(7)
The assumption of “response consistency” means that the thresholds τ i j are the same for the selfreports and the vignettes. The assumption of “vignette equivalence” implies that genuine workrelated health of the vignette person Yli* does not depend on X i ; it only depends on the vignette
description (l) and an idiosyncratic error term.8
Given these assumptions, it is clear how the vignette evaluations can be used to
separately identify β and γ (=γ 1 ,...γ 4 ) : From the vignette evaluations alone, γ , θ1 ,...θ L can be
identified (up to the usual normalization of scale and location). From the self-reports, β can then
j
The γ for j>1 will still be identified.
7
Kapteyn et al. (2007) also discuss a model in which thresholds are allowed to vary across the three domains of
work disability. They find that the results for this more general model are very similar to the results for the model
imposing equal thresholds across domains.
8
Allowing the vignette evaluations to depend upon gender of the vignette person (as was done in Kapteyn et al.,
2007) does not affect the results.
6
10
be identified in addition. Thus the vignettes can be used to solve the identification problem due
to DIF. The two-step procedure is sketched only to make intuitively clear why the model is
identified. In practice, all parameters will be estimated simultaneously by maximum likelihood.9
Adjusting for DIF is straightforward in this model once the parameters are estimated.
Define a benchmark respondent with characteristics Xi = X(B). (For example, choose one of the
countries as the benchmark country.) The DIF adjustment would now involve comparing Yi* to
the thresholds τ Bj rather than τ i j , where τ Bj is obtained in the same way as τ i j but using X(B)
instead of Xi. Thus a respondent’s work ability is computed using the benchmark scale instead of
the respondent’s own scale. This does not lead to an adjusted score for each individual
respondent (since Yi* is not observed) but it can be used to simulate adjusted distributions of Yi
for the whole population or conditional on some of the characteristics in Xi. Of course the
adjusted distribution will depend upon the chosen benchmark.
4. Responses to Vignettes in Western Europe and the U.S.
Respondents in each of the eight European countries and the U.S. were given vignettes in
three domains of work disability—pain, affect, and heart disease. In each of the three domains,
three distinct vignettes are used to describe the conditons of a hypothetical person. The actual
vignettes used are presented in Appendix B.10 Table A.1 in Appendix A compares the responses
for the pain domain, while Tables A.2 and A.3 do the same for the affect and heart disease
(CVD) domains, respectively. The numbering of vignettes corresponds to how vignettes are
9
This is more efficient than the two-step procedure. Since all error terms are independent, the likelihood
contribution is a product of univariate normal probabilities over all vignette evaluations and the self-report, which is
relatively easy to compute.
10
The selection of vignettes is the result of simulation studies with the Dutch CentERpanel, where we administered
five vignettes per domain and then estimated Chopit models that used subsets of these five. We found that the
extreme vignettes (either describing someone who is healthy or someone who is clearly too sick to work) carried
little information. We also looked at the effect of the number of vignettes per domain. Obviously, more vignettes
will lead to more accuracy, but for practical reasons the number of vignettes per domain had to be limited. The
simulation study led to the conclusion that three vignettes per domain strikes a reasonable balance between practical
feasibility and statistical accuracy.
11
presented in Appendix B. For instance, pain1 is the first pain vignette shown in Appendix B,
pain2 is the second pain vignette, etc. Although the health conditions of the persons described in
the vignettes are supposed to be the same in all countries, Tables A.1-A.3 show that there are
large differences in the evaluation frequencies between the countries.
Some differences are striking. For instance pain2 (“[Catherine] suffers from back pain
that causes stiffness in her back especially at work but is relieved with low doses of medication.
She does not have any pains other than this generalized discomfort.”) is said to provide no work
limitation by almost 25% of the American respondents, while in the European countries that
percentage is less than 4, with very little variation across European countries. Similarly for
CVD2 (“[Tom] has been diagnosed with high blood pressure. His blood pressure goes up quickly
if he feels under stress. Tom does not exercise much and is overweight.”) almost 29% of
Americans consider this to be no or only a mild work disability, while the SHARE average is
about 16%. The differences for the Affect vignettes are most pronounced. For all three vignettes
Americans are much more likely than Europeans to say that the health condition described in the
vignette does not constitute a work disability.
Table 2 summarizes the responses to the work disability vignettes in the nine countries.
The three vignettes per domain were selected to deliberately eliminate the extremes where
individuals in all countries would tend to describe the vignette person as clearly work disabled or
clearly not work disabled. As a result, the differences in the scenarios described in the vignettes
are often not large (at least among the European countries) and there is noise in the ranking both
within and across countries. Table 2 summarizes the vignette evaluations in two ways. First, the
five-point scale on work disability is collapsed into three groups, with none and mild combined
into one group and severe an extreme into another group. The percentage of respondents
12
reporting in two of these three groups (none/mild and severe/extreme) is listed for each of the
three vignettes in each of the three domains—pain, affect, and CVD.
The second way of arriving at a simple summary is contained in the rows labeled
“Average rank” under each domain. This row is derived by ranking the countries in each of the
three vignettes within a domain by which country is ‘hardest’ on the vignette persons—that is
which country has the highest percentage of responses in the mild or no work disability category
and the lowest percentage of responses in the severe or extreme category. The final row in
Table 2 labeled “Grand Average Rank” is just the average of these ranks across the three work
disability domains.
Before we concentrate on the patterns in Table 2, we first discuss some correlations
summarized in Table 3. The upper panel shows correlations between the average rankings of
countries within domains. For example, the entry 0.44 in the upper left corner is the correlation
of the average ranking of countries obtained by calculating the average ranking for vignette 1
and for vignette 2 in the pain domain. One would expect that if countries differ strongly in how
“soft” or “tough” they are in their vignette evaluations, the ranking of countries would be pretty
much the same for each vignette. Table 3 indeed shows that all correlations but one are positive,
but often the correlations are not particularly high. The correlations appear lowest for pain and
highest for CVD. Thus, in particular for pain there may be a fair amount of noise in the vignette
evaluations.
The bottom panel shows the correlation of rank averages across domains. For example,
the entry 0.69 is the correlation between the average ranking of countries for pain and for affect,
as shown in Table 3. We observe that the correlation of rank averages across domains is much
higher than within domains (the average of the three correlations in the bottom panel is 0.67,
13
while the average of the nine correlations in the top panel equals 0.40). This supports the notion
that the sometimes weak correlations in the top panel of Table 3 are at least partly the result of
noise. Aggregating across vignettes leads to a considerable reduction of noise and thus to higher
correlations in rankings of countries based on different vignettes.
Table 2 indicates several salient patterns. First, residents of the eight European countries
do not share a common view on what constitutes a work disability. For example, while a third of
the Dutch respondents state that the first CVD vignette constitutes no or only a mild work
disability, the comparable fraction for Spaniards is one-in-twenty and for Swedes one in fifty.
Yet, the variation within Europe is less striking than the difference between Europe and the US.
Comparing the last two columns, which present the US and the SHARE averages, we see that
with the exception of pain3 (“[Mark] has pain in his back and legs, and the pain is present almost
all the time. It gets worse while he is working. Although medication helps, he feels
uncomfortable when moving around, holding and lifting things at work.”) European respondents
are always more likely to call a vignette person work disabled than Americans.
Second, the ranking amongst the European countries depends to some extent on the
specific domain chosen. For example, the Italians are quite demanding (“tough”) in the affect
domain, but the Dutch are the toughest on CVD. Yet the differences in ranking across domains
are not dramatic, consistent with the earlier observed high correlations across domains. The
Americans are the toughest overall, followed by the Italians, the Belgians, the Dutch and the
French. At the other end of the spectrum, the Greeks, Swedes, and Spaniards appear most
inclined to call a health condition work limiting. The Germans are in the middle.
14
Table 2: Responses to Work Disability Vignettes
Germany Spain
PAIN
pain 1
none,mild
severe,extreme
pain 2
none,mild
severe,extreme
pain 3
none,mild
severe,extreme
Average rank
AFFECT
Affect 1
none,mild
severe,extreme
Affect 2
none,mild
severe,extreme
Affect 3
none,mild
severe,extreme
Average rank
CVD
CVD 1
none,mild
severe,extreme
CVD 2
none,mild
severe,extreme
CVD 3
none,mild
severe,extreme
Average rank
Grand Average
Rank
Greece
Italy Netherlands Sweden
France
Belgium
US
Total EU
10.9
54.1
5.5
72
12.1
61.9
13.4
56.4
6.6
64.8
6
80
10.7
48.6
7.7
60.7
10.1
55
10.1
58.7
29
16
18.2
36.9
31.6
20.4
38.1
13.9
58.9
11.2
10.4
54
38.6
8.7
47.7
10.5
77.9
3.4
32.8
18.7
7.5
70.4
4.8
1.9
72.7
8.3
7.8
68.4
4.8
14.4
51.9
3.0
4.9
59.3
5.0
11.4
51.5
6.3
5.8
68.5
3.8
7.2
59.9
4.3
4.9
74.1
4.3
7.9
64.5
23.9
27.5
23.7
30.3
20.6
35.7
39.2
17.6
29.3
24.2
34
16.5
26.9
23
35
23.5
53.3
14.5
28.7
24.6
28.2
22.7
21
30.6
22.4
33.2
35.6
16.9
18.7
30.4
14.6
52.8
26.2
24.5
22.4
25.7
55.5
12.2
26.9
25.1
54.8
4.5
4.7
39.3
17.4
7.7
43.9
15.9
7.5
57.2
10.7
3.2
74.8
6.5
5.3
9.9
57.8
7.0
52.9
5.8
4.7
67.6
4.4
3.8
80.6
3.6
1.0
52
10.8
15.9
37.5
5.7
63.8
14.9
51.9
23.7
30
33.6
27.5
1.7
88.5
16
47.7
22.2
31.9
43.3
19
16.7
43.4
14.6
41.8
4.4
65.9
5.8
77
26.2
33.7
30.5
26.6
26.2
26.1
13.6
46.9
21.3
33.1
28.6
30.5
16.6
44.5
7.2
71.3
5.5
3.8
81.8
8.0
5.1
85.9
7.7
16.1
62.4
3.2
18.2
46.7
1.5
2.3
86.3
6.7
6.9
75.8
6.2
10.8
59.6
3.7
10.7
64.3
2.5
9.1
71.1
5.0
8.0
6.7
3.1
3.9
6.7
4.9
3.9
2.6
15
Table 3: Correlations between Rankings
Correlations between rankings within domains
Pain
0.44
0.06
-0.33
Vignettes 1 and 2
Vignettes 1 and 3
Vignettes 2 and 3
Affect
0.43
0.45
0.72
CVD
0.51
0.89
0.47
Correlations of average rankings across domains
Correlation between rank averages of pain and affect
Correlation between rank averages of pain and CVD
Correlation between rank averages of affect and CVD
0.69
0.57
0.76
It is of interest to relate national norms about work disability to institutional
arrangements. Table 4 provides a very simple way of doing so. The first column reproduces the
grand average ranks from Table 2. Columns 2 and 3 show two indicators of employment
protection published by OECD (2004). Both indicators aggregate in some fashion three main
domains that make it difficult for an employer to dismiss an employee. These domains are (see
OECD, 2004, p. 65):
i) difficulty of dismissal, that is legislative provisions setting conditions under
which a dismissal is “justified” or “fair”;
ii) procedural inconveniences that the employer may face when starting the
dismissal process;
iii) notice and severance pay provisions.
Version 2 is somewhat broader than version 1.
There is a fairly strong positive correlation between the strength of employment
protection and a country’s rank in the vignette distribution. People in countries with more
employment protection are on average “softer” on work disability (i.e. more inclined to see a
16
given health condition as work limiting). Naturally, there are various alternative explanations for
this positive correlation. One would be “culture”. In a country with a tough culture, citizens are
tough on work disability and don’t find employment protection very important. This is then
reflected in laws with little protection. The US would be a case in point.
Table 4: National Norms about Work Disability and
Employment Protection
Italy
Belgium
Netherlands
France
Germany
Greece
Sweden
Spain
US
Correlation with rank
Rank
3,1
3,9
3,9
4,9
5
6,7
6,7
8
2,6
1,00
OECD Employment
protection indicator
version 1 version 2
3,1
3,1
2,2
2,5
2,1
2,3
3
2,9
2,2
2,5
2,8
2,9
2,2
2,6
3,1
3,1
0,2
0,7
0,52
0,56
5. Model Estimates for Response Scales and Self reported Work Disability
In this section, we present our parameter estimates for predicting work disability in the
nine countries. Our models incorporate a number of standard demographic covariates—age
dummies (less than 58, 58-64, 65-71, 72 or more), years of education, dummies for being female
and for currently married. In addition, a series of health indicators are included (heart problems,
lung disease, high blood pressure, diabetes, pain, arthritis, cancer, CESD score) with for the
European countries the EuroD. The benchmark country is the US. We include a full set of
interactions with an EU dummy. Furthermore we include country dummies for the separate EU
countries. Table 5 lists the means of the explanatory variables and self-reported work disability
by country.
17
Country
USA
Germany
Sweden
Netherlands
Spain
Italy
France
Greece
Belgium
Table 5A. Sample Means by Country: Demographics
female
marrlt
educyrs
age 58-64 age 65-71
0.543
0.693
13.026
0.204
0.227
0.558
0.632
13.409
0.229
0.234
0.510
0.669
10.674
0.267
0.147
0.509
0.726
11.974
0.236
0.193
0.545
0.632
6.682
0.226
0.181
0.549
0.615
7.239
0.266
0.184
0.558
0.702
9.559
0.221
0.191
0.534
0.685
9.205
0.237
0.246
0.535
0.737
10.594
0.202
0.216
age72+
0.151
0.241
0.267
0.214
0.288
0.261
0.300
0.241
0.265
Explanation: female: dummy female; marrlt: dummy married or living together (benchmark: single,
divorced, separated, or widowed); educyrs: years of education; age58-64, age65-71, age72+: dummies
age groups 58-64, 65-71 and 72 and older, respectively (benchmark: 50-57). All sample means are
weighted.
Country
USA
Germany
Sweden
Netherlands
Spain
Italy
France
Greece
Belgium
Country
USA
Germany
Sweden
Netherlands
Spain
Italy
France
Greece
Belgium
Table 5B. Sample Means by Country: Health Conditions
heart
lung
hbp
diabetes
pain
0.192
0.083
0.492
0.160
0.373
0.113
0.069
0.343
0.102
0.567
0.121
0.095
0.303
0.081
0.552
0.099
0.100
0.239
0.081
0.355
0.101
0.080
0.343
0.119
0.479
0.118
0.131
0.366
0.087
0.591
0.144
0.087
0.288
0.101
0.531
0.154
0.050
0.326
0.120
0.503
0.141
0.063
0.286
0.074
0.504
cancer
0.107
0.063
0.085
0.052
0.058
0.054
0.064
0.027
0.067
obese
0.300
0.175
0.155
0.126
0.212
0.152
0.158
0.168
0.196
cesd
1.302
0.992
0.964
0.909
1.327
1.446
1.353
1.166
1.128
arthritis
0.511
0.131
0.096
0.093
0.239
0.364
0.311
0.195
0.256
sdis
1.967
1.994
1.983
1.764
2.028
1.967
1.906
1.624
2.026
Explanation: heart, …cancer: dummies for chronic conditions based upon answers to survey questions
“has the doctor ever told you that you have ….”; heart: heart problems; lung: lung disease; hbp: high
blood pressure; pain: 1 if answer is “yes” to “do you often have pain?”; obese: BMI>30 (based upon selfreported weight and height); cesd: Center for Epidemiological Studies depression score in HRS;
0.5*EURO-D depression score in SHARE; a higher value indicates more depression related symptoms;
sdis: self-reported work disability on a scale from 1 (none) to 5 (extreme). All means are weighted.
18
The first part of the table shows large differences in years of education, with low means
in the southern European countries. There are also substantial differences in the age composition,
with, for example, relatively few 65-71 year olds in Sweden. Most chronic conditions are much
more prevalent in the US than in most European countries. This applies in particular to obesity,
high blood pressure, diabetes, arthritis and heart problems. Still, self-reported work disability in
the US seems well in line with that in the European countries. Only Greek and to a lesser extent
Dutch respondents seem to face fewer work related health impairments. The vignette evaluations
suggest that, compared to Europeans, US respondents underemphasize work related health
impairments, so that their actual work related health limitations may be larger than those
suggested by the means in Table 5B. Simulations based upon the estimates of the econometric
model will look at this in detail.11
Table A.4 – A.6 in the appendix present estimation results for the complete model,
allowing for DIF, as well as for a model without DIF. The latter model is a standard ordered
probit explaining the self-reported work disability on a five point scale. The differences between
the two models illustrate the effects of allowing for differences in response scales. The model not
allowing for response scale variation is strongly rejected by the data, as is immediately apparent
from a comparison of the log-likelihoods (at the bottom of Table A.6).
Table A.4 present the estimates of the work disability equation. The estimated
coefficients for the demographic attributes yield few surprises. Work disability increases with
age, decreases with schooling and is lower for married respondents. Having any of the health
conditions that are included in the model makes it more likely that one reports a work disability.
The interactions with the dummy for Europe show that in particular the effects of cancer and
11
The difference in wording between the work disability self-assessment in the US and Europe might also play a
role.
19
depression on work disability are stronger in the European countries than in the US. No such
effects are found in the model that does not correct for DIF. The country dummies in Table A.4
demonstrate that, once we correct for response scale differences and control for health
conditions, there are still substantial differences across European countries, with a particularly
low probability to be work disabled in Greece. (The US is the reference category but because of
the interactions with the EU dummy, the coefficients cannot be used to compare the European
countries with the US; this will be done in the simulations below). The effects are similar, but
smaller, in the model without DIF. Both the models with and without DIF show that the effect of
education or being married on work disability is less in Europe than in the US.
The significant estimates of the effect of covariates on the thresholds (Table A.5; model
with DIF only) explain why the effects of demographics and health conditions on self-reported
work disability change once we take response scale differences into account. Comparing effects
of country dummies on the first threshold across the European countries, suggests that Italians
are toughest and Spaniards softest, consistent with the raw data summarized in Table 2.
Compared to country dummies, effects of health conditions and demographics on the first
threshold are modest. More educated respondents tend to be softer, albeit less so in the EU than
in the US. Females are harder in the US, but not in the EU. Older respondents are softer in the
US, but not in the EU.
The interpretation of the estimated parameters for the remaining thresholds is not very
straightforward since these parameters reflect shifts of the second threshold relative to the first
one (cf. Equation (3)). Rather than extensively discussing these parameters we turn to some
simulations that are easier to interpret.
20
5.1 Simulation Results
Figure 2 provides a first impression of the effects of using different scales. The first
column presents the distribution of self-reported work disability for the US, using US data and
US parameters. This simply predicts sample observations for the US. Similarly the last column
predicts the distribution of work disability using EU (SHARE) data and EU parameters.
Comparing the first and last column suggests relatively minor differences in work disability
prevalence between the EU and the US. For instance the percentage saying that they have no
work disability is predicted to be about 51% in the US and 49% in the EU. On the other hand the
percentage of Americans with severe or extreme work disabilities is predicted to be about 16% in
the US and only 12% in the EU. These seemingly modest differences are the result of two
counter-acting effects: Americans use response scales that are “harder” while Europeans appear
to suffer from less work disability. This is illustrated by the remaining three columns in Figure 2.
Figure 2: Simulated work limitations according to different scenarios
100%
90%
80%
70%
Extreme
60%
Severe
50%
Moderate
40%
Mild
30%
None
20%
10%
0%
U.S.
EU data/US
EU data/US
EU data/EU
pars/US scales pars/EU scales pars/US scales
21
EU
The second column uses EU data, but US scales and US parameter estimates to predict
EU disability. The distributions in columns 1 and 2 are virtually identical, with a slight shift in
the direction of more work disability in the EU. This implies that distributional differences in
demographics and health have only a minor effect on differences in observed work disability
prevalence in the US and the European countries.
The middle column is based on EU data, US parameters of the work disability equation
and EU response scales. Compared to the first and second column, we now see a dramatic shift
in the distribution in the direction of increased work disability. The shift is due to the fact that
Europeans use softer response scales and more easily call someone disabled. It is also instructive
to make a comparison with the fifth column (both EU data and EU parameters). Since the only
difference between the third and fifth column is the use of US parameters in the work disability
equation, this comparison suggests that, for given demographics and health conditions, the risk to
face an actual work disability is substantially higher in the US than in Europe. For instance,
column 3 implies that 34% has no work limitation, while column 5 implies that 49% has no work
limitation.
Finally, a comparison of columns 4 and 5 isolates the effect of response scales on
observed work disability distributions in the European countries. When we assign US scales to
EU respondents, we observe a dramatic fall in reported work disability. For instance, the
percentage of respondents without a work disability goes up from 49% to 64%, while the
percentage with severe or extreme work limitations falls from 12% to about 10.5%. Again this is
a consequence of the fact that Americans less easily classify someone with a given impairment
as work disabled.
22
Table 6 presents simulated work disability by country. The first row in each panel
presents work disability for the US; the second row presents work disability in the country, but
using US scales and the third row presents work disability simulated using the response scale of
that country.
As was observed for the set of all European countries, using US scales has dramatic
effects on the distribution of disability within each European country. It is worth noting that
generally most of the changes take place in the categories none and mild. The percentage of
respondents in the EU country whose work disability falls in the severe/extreme range is not
affected as much. This is as one would expect; there is probably less scope for disagreement
about whether a serious health condition constitutes a work disability than whether a mild
condition should be seen as work limiting. The biggest reduction in the severe/extreme category
takes place in Sweden (from 14.6% to 7.7%) and Spain (from 14.6% to 10.5%). In other
countries the changes are much less dramatic and in some cases the adoption of the US scales
leads to an increase in the percentage of respondents with severe or extreme work disabilities.
This happens for Italy (an increase by 1.4 percentage points) and The Netherlands (an increase
by .3 percentage points).
When considering the changes in the categories none and mild we see the biggest
decrease in Spain (by 12.8 percentage points) and Sweden (12.2 percentage points). The smallest
decrease is seen in The Netherlands (2.6 percentage points) and Italy (4.6 percentage points).
We have seen earlier that the use of US scales actually increases the difference in work
disability between the US and the EU countries we are considering here. It is also of interest to
investigate if the use of vignettes reduces the differences between the EU countries. A simple
measure is the variance of the percent none/mild and of the percent severe/extreme across
23
European countries. We find that the variance of the percentage none/mild is 36.9 if we use the
own country scales and 26.5 if we use the common (US-) scale. For the percentage
severe/extreme, the use of a common scale has virtually no effect. Using the own scale the
variance is equal to 11.0 and when using the common scale the variance equals 10.8.
Table 6: Work Disability by Country
Germany
US
EU, US scales
EU
Sweden
US
EU, US scales
EU
Netherlands
US
EU, US scales
EU
Spain
US
EU, US scales
EU
Italy
US
EU, US scales
EU
France
US
EU, US scales
EU
Greece
US
EU, US scales
EU
Belgium
US
EU, US scales
EU
None
Mild
Moderate
Severe
Extreme
None/
Mild
Severe/
Extreme
51.4
61.2
45.4
19.1
17.7
23.5
13.2
11.0
18.7
9.9
6.9
9.8
6.5
3.2
2.6
70.4
78.8
68.8
16.4
10.1
12.5
51.4
68.6
52.2
19.1
15.1
19.3
13.2
8.7
14.0
9.9
5.2
10.5
6.5
2.5
4.1
70.4
83.6
71.4
16.4
7.7
14.6
51.4
69.6
53.2
19.1
15.3
29.0
13.2
8.3
11.3
9.9
4.8
4.4
6.5
2.1
2.1
70.4
84.8
82.2
16.4
6.8
6.5
51.4
68.6
45.6
19.1
12.8
22.9
13.2
8.1
16.9
9.9
5.7
11.1
6.5
4.8
3.5
70.4
81.4
68.5
16.4
10.5
14.6
51.4
60.0
47.6
19.1
15.1
22.9
13.2
10.3
16.4
9.9
7.8
7.7
6.5
6.7
5.4
70.4
75.1
70.5
16.4
14.5
13.1
51.4
65.8
52.0
19.1
15.4
20.6
13.2
9.2
16.7
9.9
6.0
8.4
6.5
3.6
2.2
70.4
81.2
72.7
16.4
9.6
10.6
51.4
81.5
68.9
19.1
9.5
15.1
13.2
4.9
9.4
9.9
2.7
4.5
6.5
1.4
2.1
70.4
91.1
84.0
16.4
4.1
6.6
51.4
59.0
43.4
19.1
17.0
26.7
13.2
11.3
16.7
9.9
7.9
8.9
6.5
4.8
4.4
70.4
76.0
70.1
16.4
12.7
13.2
24
Next we consider a number of socio-economic categories. Table 7 shows results by age.
The scale effects appear to be fairly uniform by age. For instance the effect of going from
European scales to US scales on the percentage severe/extreme varies from 1.2 in the 58-64
category to 1.8 in the 72+ category. Similarly the effect on the percentage none/mild varies from
-7.5 in the 58-64 category to -9.9 in the 72+ category.
Table 7: Work Disability by Age in the EU and the US
US
50-57
EU
US,
sc.
EU
US
58-64
EU
US,
sc.
EU
US
65-71
EU
US,
sc.
EU
US
72plus
EU
EU
US,
sc.
US
All
EU
US,
sc.
EU
None
59.8
75.5
59.4
49.8
71.4
55.5
48.4
61.4
45.9
34.7
48.2
32.9
51.4
64.5
48.8
Mild
17.7
12.7
20.8
19.6
14.2
22.5
19.5
16.2
23.7
21.2
18.6
24.0
19.1
15.3
22.6
Moderate
10.7
6.4
12.5
13.3
7.7
14.0
14.2
10.6
17.4
18.4
14.3
22.5
13.2
9.6
16.4
Severe
7.3
3.7
5.5
10.3
4.6
6.1
10.8
7.3
9.3
14.9
10.5
14.4
9.9
6.4
8.7
Extreme
4.5
1.9
1.8
7.0
2.2
1.9
7.1
4.6
3.7
10.9
8.4
6.3
6.5
4.2
3.4
None/Mild
77.5
88.1
80.3
69.4
85.6
78.1
67.9
77.6
69.6
55.9
66.8
56.9
70.4
79.8
71.4
Sev/Extr.
11.8
5.5
7.2
17.3
6.8
8.0
17.9
11.8
13.0
25.8
18.9
20.7
16.4
10.6
12.1
6. Work Disability, Work, and Response Scales
We now extend the model presented in Section 3 by adding an equation explaining
employment status, while allowing for justification bias. Justification bias has been introduced
by Bound (1991) as a possible effect of non-employment on self reported work disability, where
respondents may exaggerate their work limitations to justify that they don’t work. The empirical
evidence for the existence of justification bias appears to be mixed (see for instance Kreider and
25
Pepper (2007) or Jones (2007) and references therein). The use of vignettes provides a novel
approach to the estimation of justification bias.
As before, Yi denotes the answer of respondent i to the five-point scale self-assessed work
disability question while the answers to the work limitations of the vignette persons on the same
five point scale are denoted by Yli where each respondent i evaluates L vignettes l=1,…,L.
For employment status, we use a binary variable E, with Ei=1 if respondent i does some
paid work, and Ei=0 otherwise. In principle we could distinguish more than two categories here
(unemployed, on disability benefits, retired, homemaker, …) but the numbers of observations in
some of these employment states are quite small in some countries, making estimation of a richer
model difficult.
The equation for self-reported work disability is the same as before, cf. (1)-(2). To keep
the model manageable by limiting the number of parameters to be estimated and to facilitate
interpretation of the results, we replace the threshold equations (3)-(4) by:
τ i0 = −∞, τ i5 = ∞, τ i1 = X iγ + δ Ei + ui , τ i j = τ i j −1 + γ j , j = 2, 3, 4
(8)
ui ∼ N (0, σ u2 ), independent of ε i and X i
(9)
This is a more parsimonious specification than the one used in (3). All threshold shifts are the
same. The error term ui again reflects unobserved heterogeneity in thresholds. The effect of
employment status Ei on the response scales reflects justification bias: depending on their
employment status, respondents may use different response scales. The advantage of the more
parsimonious specification is that in this model, justification bias is captured by a single
parameter δ (which will be allowed to be different for US and European respondents in the
empirical work) instead of a set of parameters affecting several thresholds differently. We expect
26
a positive estimate of δ if employed respondents are less likely to evaluate a given (vignette)
person as work disabled than non-employed respondents.
The evaluations of vignettes l=1,…,L are also modeled as before, cf. (5)-(7). Employment
status is modeled using a probit-like equation:
Ei* = X iϕ + π Yi * + ηi
ηi ∼ N (0,1), ηi independent of X i , ui , ε ri , ε1i ,..., ε Li
(10)
Ei = 1 if Ei* > 0; Ei = 0 if Ei* ≤ 0
The coefficient π represents the effect of work disability on employment status, which is the
main coefficient of interest in many studies of the effect of work related health on employment
(cf., e.g., Kreider and Pepper, 2007).
In this model, error terms are all independent of each other, and no exclusion restrictions
are needed for identification. The intuitive argument for this is as follows: equations (5) and (8)
are identified from the vignette evaluations alone, since employment status is exogenous to the
errors in (5) and (8). Equation (1) is then identified as well, just like in the model of Section 3.
Since the error term ε i in (1) is independent of the error term in (10), the variation in Yi * induced
by the variation in ε i is sufficient to identify equation (10) (although estimation of (10) is
complicated by the categorical nature of Yi ).
As a sensitivity check, we will also present results for an over identified version of the
model where we impose the exclusion restriction in the employment equation that health
conditions affect employment status through work disability only and not directly (maintaining
the assumption that all errors are independent of each other).
Figure 3 presents the model including work and justification bias in graphical form.
Actual work disability affects employment status (the parameter π in (10)), but we assume there
27
is no reverse effect of employment status on actual work disability, in line with, for example, the
results of a recent panel data study of Böckerman and Ilmakkunnas (2009). We could in principle
allow for a reverse effect if appropriate exclusion restrictions were available. In our model,
employment status affects reported work disability through the justification bias effect on
response scales (the parameter δ ).
The figure does not show the vignette evaluations. We maintain the assumptions
underlying the vignette approach discussed in section 3, in particular vignette equivalence and
response consistency. The first now also implies that workers and non-workers do not interpret
the genuine work related health of the vignette persons in systematically different ways, and we
see no reason not to find this assumption plausible. Response consistency now also implies that
justification bias plays the same role in vignette evaluations as in self-assessments. This is an
identifying assumption that cannot be tested with the current data. Thus if non-workers evaluate
themselves as more work limited than workers with the same actual work related health, they are
assumed to also do this with the vignette persons. One might argue that they will be less inclined
to exaggerate the vignette person’s health problems than their own impairment, which would
imply that the effect of employment status on the vignette evaluations is an underestimate of
justification bias in self-assessments. Since our estimate of justification bias is driven by the
effect on vignette evaluations, this would imply that with our method, we get a lower bound on
the “true” justification bias in self-assessments.
28
Figure 3. Work, work disability, response scales, and justification bias
7. Model Estimates for Work Disability, Work, and Threshold Equations
Tables A.7-A.11 present the model estimates with and without DIF. Comparing Table
A.7 with the estimates for the disability equation in Table A.4 shows that generally the
coefficients are of a similar order of magnitude and have the same sign. One notable difference is
the coefficient on gender, which is significantly negative in Table A.7, and which was
completely insignificant in Table A.4.
The results for the threshold equation given in Table A.8 cannot be compared directly to
Table A.5, as the specification is much simpler. We will use simulations (see below) to see if the
simpler specification leads to very different outcomes. The country dummies (with the US as the
reference category) appear to be smaller, although Greece maintains its status as somewhat of an
29
outlier. The parameter of primary interest in Table A.8 is the coefficient on the dummy Work.
The coefficient is significantly positive in the US, while the interaction with the EU dummy
makes clear that the effect of employment status on the thresholds is essentially zero in the EU.
In other words, we find significant justification bias in the U.S. and no evidence of justification
bias in the EU. It suggests that not working is much more accepted in the EU (for whatever
reason) than in the US. Even if our estimate of justification bias is an underestimate of the true
justification bias since it does not affect vignette evaluations as much as self-assessments (see the
end of the previous section), we still believe this conclusion is justified, since we see no reason
why the violation of response consistency would apply in Europe but not in the US. It might
mean that justification bias is present in Europe as well as the US, but then it is still much bigger
in the US than in Europe.
The estimates of the vignette equation are presented in Table A9. The estimated
coefficients on the vignette dummies are similar to those in Table A.6, although they do not
show exactly the same ranking across vignettes. Still, the correlation between the dummy
estimates in the two tables is 0.97.
The employment equation estimates are shown in Table A.10. Since the model without
DIF is soundly rejected, we concentrate on the estimates for the model with DIF. Since (genuine)
work disability is controlled for, the effects of the health dummies indicate the effect of a
specific condition on employment keeping work disability constant. The negative effects of
mental health (measured through the CESD score) and, in the US, diabetes on employment
therefore imply that the negative effects of these health conditions on employment are larger than
their effect through work disability would suggest. On the other hand, the effect of obesity on
employment in the US is smaller than its effect through work disability. The insignificant effects
30
of other health conditions simply mean that the effects of conditions are well captured by the
effect of work disability. The estimated age dummies show the expected pattern that employment
falls with age. This pattern is much stronger in the EU than in the US. Similarly, females are less
likely to be employed than males and also this difference is stronger in the EU than in the US.
The parameter of main interest in the employment equation is the coefficient of work
disability. It is more than twice as large in absolute terms in the US than in the EU (-.464 in the
US and -.192 in the EU). This result remains essentially unchanged when we specify the
employment equations more parsimoniously by omitting the health conditions (Table A.11). One
possible interpretation of this is that individuals in the US are more likely to work for pay and
that health is one of the main impediments for doing so. In the EU on the other hand, also
individuals in relatively good health are often not working, for instance as a result of more
generous income replacement schemes. Below we will present simulations that shed more light
on this (and other) interpretation(s).
Comparing the coefficient of work disability with the estimate of the same coefficient in
a model without DIF shows the effect of controlling for justification bias when estimating the
effect of work disability on employment. For the European countries, there is virtually no
difference between the DIF corrected and not DIF corrected estimates in Table A.10 (-.192 and
-.197, respectively), which corresponds to the finding that there is no evidence of justification
bias for these countries. For the US the difference is larger (-.464 versus -.516) and, as expected,
we find that the size of the effect of work disability on employment is overestimated if
justification bias is not corrected for—part of the negative correlation between work disability
and employment is due to the fact that the non-employed tend to over report their work
disability. But the size of the bias is modest. (But then again, our estimate might be seen as a
31
lower bound on the bias, if justification bias plays less of a role in vignette evaluations than in
self-assessments; see the discussion at the end of Section 6).
7.1 Simulation Results
Figure 4 is analogous to Figure 2 above, but now based on the extended model including
an employment equation and allowing for justification bias. Compared to the model underlying
Figure 2, the current model has one extra equation (the employment equation) so that there are
more possible combinations of data and parameters that can be used in simulating
counterfactuals. The first three bars in Figure 4 are most comparable to those in Figure 2, while
the last two bars in Figure 4 correspond to the last two bars in Figure 2. Qualitatively the patterns
appear to be similar. If we move from US data to EU data (but retaining US scales and
parameters), i.e. from the first column to the second, we see again a slight shift in the direction of
more work disability in the EU than in the US.
When we move from the second to the third column (EU data, US parameters, EU scales)
we see an increase in work disability, but the shift is not as dramatic as in Figure 2. Finally, a
comparison of the last two columns shows again that if we move from EU data and EU
parameters, but US scales to all data, parameters and scales European, then work disability
increases substantially (but not as much as in Figure 2). This illustrates once again the difference
in response scales used in the EU countries and in the US. The reason that the changes are less
dramatic than in Figure 2 is probably the fact that we have imposed that all threshold shifts are
parallel to obtain a parsimonious model (see Section 6).
32
Figure 4: Simulated work limitations; model including employment equation
100%
90%
80%
70%
60%
50%
40%
30%
20%
10%
0%
Extreme
Severe
Moderate
Mild
None
US
EU-data EU-data EU-data EU-data EU-data EU-data EU-data
US-dis US-dis US-dis US-dis EU-dis EU-dis EU-dis
US-work US-work EU-work EU-work US-work US-work EU-work
USEUUSEUUSEUUSscale
scale
scale
scale
scale
scale
scale
EU
Comparing columns 4 (EU-data, US-disability equation, EU-work equation, US-scale)
and 8 (EU-data, EU-disability equation, EU-work equation, US-scale) shows a substantial fall in
the prevalence of work disability. Since the only difference between the two columns is the
difference in disability parameters, it illustrates once more the finding that Europeans are less
work disabled than the Americans, although the reported work disability levels are similar.
The only difference between columns 6 (EU-data, EU-disability equation, US-work
equation, US-scale) and 8 (EU-data, EU-disability equation, EU-work equation, US-scale) is that
in the former the US employment parameters have been used, while in the later we use the EU
employment parameters. These differences in parameters appear to have minimal impact on the
distribution of observed work disability, even though observed work disability does depend on
employment status through its (justification bias) effect on the response scales.
Table 8 and Figure 5 show the percentage of people working for each of the five levels of
work disability (the last column of Table 8 is discussed below). As could be expected on the
33
basis of the raw data, a comparison of the first and last row of Table 8 shows that Europeans
work less than Americans for any work disability category.
Table 8: Percent Working by Disability Category
US
EU-data US-dis US-work US-scale
EU-data US-dis US-work EU-scale
EU-data US-dis EU-work US-scale
EU-data US-dis EU-work EU-scale
EU-data EU-dis US-work US-scale
EU-data EU-dis US-work EU-scale
EU-data EU-dis EU-work US-scale
EU
None
66.6%
59.5%
61.8%
34.4%
35.3%
59.8%
61.8%
33.6%
34.5%
Mild
45.8%
36.5%
40.5%
22.3%
24.4%
35.3%
39.3%
20.7%
22.9%
Moderate
34.3%
25.0%
28.6%
16.7%
18.8%
24.3%
27.8%
15.4%
17.4%
Severe
23.5%
15.1%
18.0%
11.5%
13.4%
15.0%
17.7%
10.8%
12.4%
Extreme Correlations
12.2%
0.32
6.3%
0.34
7.8%
0.33
6.1%
0.20
7.4%
0.19
6.7%
0.33
8.2%
0.33
6.1%
0.20
7.3%
0.19
A comparison across the rows of Table 8 (or the columns of Figure 5) shows that a major source
of differences lies in the parameters of the employment equation. Whenever we simulate labor
market outcomes using the US employment parameters, we find considerably higher
employment rates than when we use EU parameters. This is particularly true for the less severe
work disability categories, in line with the larger effect of work disability on employment in the
US than in Europe. Once we move to extreme work disabilities the differences become small;
with an extreme work disability neither Americans nor Europeans are likely to work.
Figure 6 shows total employment rates in the US and the EU. Consistent with Table 8
and Figure 5, we see that employment is about twice as large when a US employment equation is
used than when we use the EU parameters in the employment equation. Whether the EU or the
US disability parameters are used has very little impact.
34
Figure 5: Percent working by disability category
70.0%
60.0%
None
50.0%
Mild
40.0%
Moderate
30.0%
Severe
20.0%
Extreme
10.0%
0.0%
US
EU-data
US-dis
US-work
US-scale
EU-data
US-dis
US-work
EU-scale
EU-data
US-dis
EU-work
US-scale
EU-data
US-dis
EU-work
EU-scale
EU-data
EU-dis
US-work
US-scale
EU-data
EU-dis
US-work
EU-scale
EU-data
EU-dis
EU-work
US-scale
EU
Figure 6: Employment rates
Employment rate
60.0%
50.0%
40.0%
30.0%
20.0%
Employment rate
10.0%
0.0%
US
EU-data EU-data EU-data EU-data EU-data EU-data EU-data
US-dis US-dis US-dis US-dis EU-dis EU-dis EU-dis
EUUSUSEUUSUSEUwork
work
work
work
work
work
work
EUUSEUUSEUUSUSscale
scale
scale
scale
scale
scale
scale
EU
Finally, we return to the issue of justification bias. The last column of Table 8 shows
simulated correlations between the binary outcomes employment (employment=1 and nonemployment=0) and not having a work limiting health condition (none=1, mild or worse=0).
Other things equal, one would expect a stronger correlation between work and self-reported
35
disability if US scales are used, because of justification bias, which is significant in the US and
absent in the EU. We see a small effect of justification bias. For instance, comparing the second
and third row we see the correlation decrease from .34 to .33. Other comparisons show similar
small effects. However using the US employment parameters rather than the EU ones has a much
bigger effect. This confirms once again the much weaker relation between disability and
employment in the EU than in the US.
8. Conclusions
We have provided several pieces of evidence on differences in work disability between
older Europeans and Americans. The descriptive vignette analyses suggest that opinions about
what constitutes a health related work limitation are related to the generosity of earnings
replacement schemes and employment protection in different countries. We also find that the
differences in self-reported disabilities across European countries fall substantially when we use
common reporting scales (i.e. if we use vignette corrections to make scales comparable).
Consistent with earlier research, we find that Americans are less likely to call themselves work
disabled than Europeans.
The extended model with employment included, suggests that justification bias plays a
role in the US, but not in the EU. Americans use health as a justification for not working,
whereas Europeans don’t feel the need to do so. The effect of health limitations on employment
is about twice as large in the US than in the EU. Our simulations suggest that this mainly reflects
the fact that in the EU even healthy individuals are less likely to be working than in the US.
36
References
Böckerman, P., and P. Ilmakunnas (2009), “Unemployment and self-assessed health: evidence
from panel data,” Health Economics, 18(2), 161-179.
Boersch-Supan, A. (2007), “Work Disability, Health, and Incentive Effects,” Working Paper,
MEA, University of Mannheim.
Bound, J. (1991), “Self-reported versus objective measures of health in retirement models,”
Journal of Human Resources, 26(1), 106-138.
Eurostat. (2001),.Disability and Social Participation in Europe, Luxembourg: Office for Official
Publications of the European Communities.
Croda, E. and J. Skinner (2009). “Disability Insurance and Health in Europe and the US. April,”
Paper presented at Netspar 2009 Annual Conference, The Hague,
http://www.netspar.nl/events/2009/annual/paperskinner.pdf.
Heeringa S.G., and J. Connor (1995), “Technical Description of the Health and Retirement Study
Sample Design,” Online version at www.hrsonline.isr.umich.edu; originally published as
HRS/AHEAD Documentation Report DR-002.
Jones, M.K. (2007), “How accurate are estimates of the impact of disability on labour market
participation? Evidence from the Health Survey for England,” Swansea University, School of
Business, Working Paper.
Kapteyn, A., J.P. Smith, and A. van Soest (2007), “Vignettes and self-reported work disability in
the US and the Netherlands,” American Economic Review, 97(1), 461-473.
Kapteyn, A., J.P. Smith, A. van Soest, and J. Banks (2009), “Labor Market Status and
Transitions during the Pre-Retirement Years: Learning from International Differences,”
37
forthcoming in D. Wise (ed.), Research Findings in the Economics of Aging, University of
Chicago Press.
King, G., C. Murray, J. Salomon, and A.Tandon (2004),“Enhancing the Validity and Crosscultural Comparability of Measurement in Survey Research,” American Political Science
Review, 98(1), 567-583.
Kreider, B. and J.V. Pepper (2007), “Disability and Employment: Reevaluating the Evidence in
Light of Reporting Errors,” Journal of the American Statistical Association, 102(478), 432441.
OECD (2004), OECD Employment Outlook, OECD, Paris.
Van Soest, A., L. Delaney, C. Harmon, A. Kapteyn, and J.P. Smith (2007), “Validating the Use
of Vignettes for Subjective Threshold Scales,” RAND, Labor and Population Working Paper,
http://www.rand.org/pubs/working_papers/WR501/
38
APPENDIX A. Tables
Table A.1: Responses to Pain Work Disability Vignettes
pain 1
Germany Spain Greece Italy Netherlands Sweden France Belgium Total EU
none
1.96 0.65
0.41 4.03
2.12
1.51 1.56
1.28
2.05
mild
8.91
4.8 11.65 9.33
4.48
4.5 9.14
6.41
8.02
moderate
35.08 22.56
26 30.27
28.55 13.98 40.68 31.64
31.25
severe
47.77 62.2 41.12 43.47
36.49 53.16 43.83
42
47.45
extreme
6.29 9.79 20.82 12.89
28.35 26.85 4.78 18.67
11.24
Total
100 100
100
100
100
100
100
100
100
US
1.53
8.54
34.99
48.32
6.63
100
pain 2
Germany Spain Greece Italy Netherlands Sweden France Belgium Total EU
none
2.28 1.81
3.4 6.58
2.63
0.94 5.29
2.9
3.72
mild
26.77 16.38 28.18 31.47
56.23
9.48 33.31 44.83
29.04
moderate
54.94 44.95 48.07 48.06
29.92 35.53 52.75 41.81
48.51
severe
13.85 34.78 17.22 11.73
10.72 47.59 7.67
9.59
16.71
extreme
2.15 2.09
3.13 2.16
0.5
6.46 0.98
0.87
2.02
Total
100 100
100
100
100
100
100
100
100
US
24.97
52.9
18.78
2.25
1.11
100
pain 3
Germany Spain Greece Italy Netherlands Sweden France Belgium Total EU
none
1.22
0
0.41 2.62
1.47
0.45 1.15
1.35
1.28
mild
6.29 1.94
7.39 11.78
3.46 10.93
4.7
5.86
6.62
moderate
22.06 25.41 23.77 33.73
35.78 37.14 25.7 32.88
27.61
severe
57.57 63.15 48.03 34.96
36.75 43.78 61.77 47.86
51.59
extreme
12.85
9.5 20.41 16.91
22.53
7.69 6.68 12.05
12.9
Total
100 100
100
100
100
100
100
100
100
US
1.16
3.72
21.06
50.43
23.62
100
39
Table A.2: Responses to Affect Disability Vignettes
Affect 1
none
mild
moderate
severe
extreme
Total
Germany Spain Greece Italy Netherlands Sweden France Belgium Total EU
3.78
1.12
3.78 6.27
3.94
5.32 5.28
5.54
4.3
20.07 22.62 16.77 32.91
25.37 28.69 21.59 29.44
24.4
48.64 45.91
43.7 43.25
46.54 49.47 50.15 41.55
46.68
24.78 28.72 25.68 15.33
20.26 15.49 21.7 22.56
22.12
2.71
1.62 10.07 2.25
3.89
1.03 1.28
0.91
2.5
100
100
100
100
100
100
100
100
100
US
22.12
31.15
32.21
13.26
1.26
100
Affect 2
none
mild
moderate
severe
extreme
Total
Germany Spain Greece Italy Netherlands Sweden France Belgium Total EU
3.38
1.88
2.93 5.93
2.41
1.41 4.84
3.66
3.82
24.81 19.08 19.48 29.67
16.26 13.19 21.35 18.73
23.05
49.11 48.43
44.4 47.51
50.92 32.57 49.35 51.87
48.04
21.62 27.21 27.15 14.86
25.29 45.02 22.52 22.21
22.48
1.08
3.4
6.04 2.03
5.13
7.8 1.94
3.52
2.6
100
100
100
100
100
100
100
100
100
US
18.33
37.12
32.35
10.74
1.46
100
Affect 3
none
mild
moderate
severe
extreme
Total
Germany Spain Greece Italy Netherlands Sweden France Belgium Total EU
8.58
3.41 10.81 8.87
5.9
0.91 13.09 10.87
8.33
46.2
35.9 33.06 48.36
68.9
8.94 39.77 56.78
43.66
40.69 43.28 40.22 32.11
18.69 32.32 41.38 27.99
37.23
4.53 16.98 11.87 8.96
6.51 47.38 5.02
4.15
9.61
0
0.44
4.04
1.7
0 10.45 0.74
0.22
1.16
100
100
100
100
100
100
100
100
100
US
39.54
41.02
15.86
2.71
0.87
100
40
Table A.3: Responses to Heart Disease Work Disability Vignettes
CVD1
none
mild
moderate
severe
extreme
Total
Germany Spain Greece Italy Netherlands Sweden France Belgium Total EU
US
3.1 0.63 1.41 4.08
2.47
0.15 4.11
3.3
2.88 7.84
12.78
5.1 13.44 19.61
31.1
1.52 11.88
18.86
13.85 35.41
46.57 30.52 33.26 46.33
38.96
9.78 36.27
45.97
39.82 37.76
34.47 57.95 40.42 23.96
22.37 60.56 43.69
26.32
37.6 16.12
3.08 5.81 11.46 6.02
5.09 27.99 4.06
5.55
5.85 2.87
100
100
100 100
100
100
100
100
100
100
CVD2
none
mild
moderate
severe
extreme
missing
Total
Germany Spain Greece
3.31
0.3
0.6
11.27 4.09 5.16
43.46 29.74 17.2
39.15 54.3 49.3
2.45 11.57 27.74
0.36
0
0
100
100
100
CVD3
none
mild
moderate
severe
extreme
Total
Germany
3.99
3.2
21.55
53.39
17.87
100
Italy Netherlands Sweden France Belgium Total EU
US
4.47
3.76
1.95 3.62
2.46
2.98 8.16
20.73
26.7
24.3 9.95
18.89
13.47 20.46
38.65
42.9 47.68 39.5
45.51
38.5 40.88
24.12
21.15 22.22 41.95
30.44
37.01 26.93
8.31
5.49
3.85 4.98
2.7
7.08 3.57
3.72
0
0
0
0
0.96
100
100
100
100
100
100
100
Spain Greece Italy Netherlands Sweden France Belgium Total EU
US
0.45 1.86 5.31
3.04
0.15 2.14
3.83
3.12
3.1
3.39 3.25 10.75
15.15
2.12 4.77
6.93
5.99 7.55
14.35 8.96 21.55
35.13
11.4 17.34
29.66
19.8 24.99
57.53 33.63 36.62
32.8 54.02 55.52
42.04
48.11 48.61
24.28 52.29 25.77
13.89 32.32 20.23
17.54
22.99 15.74
100
100 100
100
100
100
100
100
100
41
Table A.4: Estimates Model Section 3: Work Disability (Eq. (1))
Constant
Female
Married/LT
Educyrs
Heart prob
Lung dis
High blood
Diabetes
Pain
Arthritis
Cancer
Cesd score
Obese
Age 58-64
Age 65-71
Age 72+
Model without DIF
coeff.
st. error
-0.135
0.107
-0.103*
0.043
-0.179*
0.045
-0.046*
0.007
0.487*
0.048
0.421*
0.068
0.145*
0.044
0.325*
0.052
0.431*
0.045
0.352*
0.044
0.142*
0.061
0.171*
0.011
0.163*
0.047
0.154*
0.058
0.130*
0.056
0.461*
0.064
Model with DIF
coeff.
st. error
-0.457*
0.131
-0.020
0.049
-0.136*
0.052
-0.043*
0.008
0.437*
0.054
0.391*
0.077
0.114*
0.049
0.241*
0.061
0.384*
0.05
0.342*
0.049
0.185*
0.069
0.139*
0.013
0.130*
0.053
0.141*
0.065
0.112+
0.063
0.372*
0.072
Female*EU
Marr/LT*EU
Educyrs*EU
Heart*EU
Lung*EU
Highbl*EU
Diabetes*EU
Pain*EU
Arthr*EU
Cancer*EU
Cesd*EU
Obese*EU
Age58-64*EU
Age65-71*EU
Age72+*EU
-0.014
0.113+
0.041*
-0.032
-0.139
-0.068
-0.115
0.005
-0.019
0.260*
0.114*
0.008
-0.015
0.132+
0.012
-0.090
0.084
0.036*
0.013
-0.044
-0.047
-0.011
0.067
0.008
0.235*
0.148*
0.035
-0.002
0.170*
0.116
0.058
0.063
0.007
0.070
0.090
0.059
0.079
0.060
0.065
0.098
0.020
0.067
0.077
0.077
0.083
0.062
0.068
0.008
0.075
0.097
0.063
0.087
0.064
0.069
0.102
0.021
0.072
0.082
0.083
0.090
Germany
-0.497*
0.139
-0.913*
0.153
Sweden
-0.600*
0.135
-1.130*
0.152
Netherland
-0.609*
0.137
-0.993*
0.151
Spain
-0.642*
0.134
-1.279*
0.149
Italy
-0.790*
0.137
-1.112*
0.152
France
-0.840*
0.130
-1.244*
0.145
Greece
-1.254*
0.133
-1.767*
0.149
Belgium
-0.535*
0.135
-0.891*
0.150
*,+: significant at two-sided 5% and 10% level, respectively
42
Table A.5: Estimates Model Section 3: Thresholds (Eq. (3))
Threshold 1
parameter
standard
error
log(threshold 2 threshold 1)
parameter
standard
error
log(threshold 3 threshold 2)
parameter
standard
error
log(threshold 4 threshold 3)
parameter
standard
error
0
0.070*
-0.011
-0.012*
-0.023
-0.004
-0.014
-0.068*
-0.053*
-0.017
0.044+
-0.013*
-0.008
-0.021
-0.037+
-0.108*
0
0.016
0.017
0.003
0.020
0.030
0.017
0.019
0.018
0.017
0.025
0.005
0.017
0.024
0.021
0.024
-0.726*
0.040*
0.071*
0.017*
0.017
-0.030
-0.008
-0.002
0.020
0.021
0.014
-0.008
-0.021
0.016
-0.010
0.048+
0.048
0.018
0.020
0.003
0.022
0.032
0.018
0.024
0.019
0.019
0.026
0.005
0.020
0.025
0.024
0.028
-0.836*
0.012
0.062*
0.019*
-0.008
0.026
-0.003
0.027
0.036+
0.000
0.014
-0.014*
-0.022
0.019
0.039
0.126*
0.049
0.019
0.020
0.003
0.022
0.031
0.019
0.024
0.019
0.020
0.028
0.005
0.021
0.027
0.025
0.028
-0.584*
-0.072*
0.043*
0.024*
0.004
0.001
-0.011
-0.055*
0.002
0.030
0.051+
-0.016*
-0.018
0.050+
0.051+
0.080*
0.052
0.020
0.021
0.003
0.023
0.035
0.020
0.026
0.022
0.021
0.031
0.005
0.022
0.028
0.027
0.030
Interaction EU
Female
-0.077*
Married/LT
0.021
Educyrs
0.007*
Heart prob
0.012
Lung dis
0.052
High blood
-0.013
Diabetes
0.156*
Pain
0.014
Arthritis
0.023
Cancer
-0.085*
Cesd score
-0.005
Obese
0.061*
Age 58-64
0.033
Age 65-71
0.061*
Age 72+
0.093*
0.022
0.024
0.003
0.030
0.039
0.023
0.029
0.025
0.027
0.038
0.009
0.027
0.031
0.029
0.033
0.009
-0.061*
-0.013*
-0.041
0.045
0.025
-0.098*
0.013
-0.023
0.043
-0.001
-0.055+
0.020
0.048
-0.012
0.025
0.028
0.003
0.035
0.044
0.026
0.038
0.026
0.029
0.043
0.010
0.030
0.033
0.034
0.038
-0.039+
-0.060*
-0.018*
0.005
-0.047
0.046*
-0.065*
0.023
-0.039
0.007
0.026*
-0.014
0.012
-0.023
-0.066*
0.022
0.025
0.003
0.029
0.037
0.023
0.032
0.023
0.025
0.037
0.007
0.027
0.031
0.030
0.033
0.047+
-0.042
-0.022*
-0.009
-0.017
-0.012
0.047
-0.021
-0.026
-0.004
0.027*
-0.058*
-0.001
-0.016
0.017
0.024
0.026
0.003
0.031
0.043
0.025
0.034
0.026
0.028
0.043
0.008
0.028
0.034
0.033
0.035
Country Dummies
Germany
-0.639*
Sweden
-0.657*
Netherland
-0.659*
Spain
-0.915*
Italy
-0.527*
France
-0.575*
Greece
-0.662*
Belgium
-0.630*
0.054
0.062
0.051
0.059
0.051
0.051
0.057
0.054
0.338*
0.180*
0.665*
0.447*
0.441*
0.289*
0.238*
0.523*
0.062
0.069
0.058
0.063
0.058
0.058
0.060
0.059
0.586*
0.285*
0.478*
0.550*
0.558*
0.589*
0.395*
0.521*
0.055
0.057
0.055
0.055
0.054
0.053
0.054
0.054
0.560*
0.419*
0.134*
0.637*
0.160*
0.551*
0.217*
0.313*
0.059
0.058
0.059
0.056
0.059
0.057
0.057
0.058
US
Constant
Female
Married/LT
Educyrs
Heart prob
Lung dis
High blood
Diabetes
Pain
Arthritis
Cancer
Cesd score
Obese
Age 58-64
Age 65-71
Age 72+
*,+: significant at two-sided 5% and 10% level, respectively
43
Table A.6: Estimates Model Section 3: Other Parameters & Log Likelihood
d vig pain 1
d vig pain 2
d vig pain 3
d vig aff 1
d vig aff 2
d vig aff 3
d vig cvd 1
d vig cvd 2
d vig cvd 3
Model without DIF
parameter
standard error
Vignette dummies
1.621*
0.024
0.747*
0.015
1.779*
0.025
0.928*
0.017
0.982*
0.018
0.563*
0.014
1.217*
0.019
1.309*
0.020
1.798*
0.025
Model with DIF
parameter
standard error
1.003*
0.202*
1.145*
0.372*
0.424*
0.019
0.641*
0.720*
1.167*
0.044
0.042
0.045
0.042
0.043
0.042
0.043
0.043
0.045
Standard deviations of errors / unobserved heterogeneity
1
(normalization)
sig selfreport
1
(normalization)
sig threshold
0.423*
0.007
sig vignette
0.728*
0.009
0.513*
0.007
const thrh
const thrh
const thrh
const thrh
Log likelihood
Threshold parameters
0
(normalization)
-0.299*
0.013
-0.290*
0.013
-0.109*
0.013
See Table A.5
-86369.50
-96584.35
*= significant at two-sided 5% level.
44
Table A.7: Estimates Model Section 6: Work Disability Equation (Eq. (1))
constant
Female
Married/LT
Educyrs
Heart prob
Lung dis
High blood
Diabetes
Pain
Arthritis
Cancer
Cesd score
Obese
Age 58-64
Age 65-71
Age 72+
Model without DIF
coeff.
St. error
-0.126
0.107
-0.106*
0.043
-0.175*
0.045
-0.046*
0.007
0.484*
0.048
0.423*
0.068
0.139*
0.044
0.321*
0.052
0.426*
0.045
0.354*
0.044
0.142*
0.061
0.170*
0.011
0.167*
0.047
0.157*
0.058
0.138*
0.056
0.463*
0.064
Model with DIF
coeff.
st. error
-0.189
0.118
-0.017
0.045
-0.118* 0.047
-0.039* 0.007
0.463* 0.050
0.421* 0.071
0.112* 0.045
0.256* 0.056
0.411* 0.047
0.364* 0.046
0.206* 0.064
0.148* 0.011
0.137* 0.049
0.166* 0.061
0.135* 0.059
0.426* 0.068
Female*EU
Marr/LT*EU
Educyrs*EU
Heart*EU
Lung*EU
Highbl*EU
Diabetes*EU
Pain*EU
Arthr*EU
Cancer*EU
Cesd*EU
Obese*EU
Age58-64*EU
Age65-71*EU
Age72+*EU
-0.007
0.110+
0.042*
-0.027
-0.141
-0.065
-0.110
0.006
-0.019
0.262*
0.114*
0.004
-0.018
0.122
0.010
0.058
0.063
0.007
0.070
0.090
0.059
0.079
0.060
0.065
0.098
0.020
0.067
0.077
0.077
0.083
-0.075
0.078
0.034*
-0.033
-0.098
-0.028
-0.076
0.048
-0.038
0.209*
0.126*
-0.020
0.004
0.161*
0.087
0.060
0.065
0.007
0.073
0.094
0.061
0.083
0.062
0.067
0.102
0.020
0.069
0.079
0.080
0.087
Germany
Sweden
Netherland
Spain
Italy
France
Greece
Belgium
-0.503*
-0.612*
-0.615*
-0.646*
-0.798*
-0.848*
-1.260*
-0.542*
0.139
0.136
0.137
0.134
0.137
0.13
0.133
0.135
-0.689*
-1.066*
-0.699*
-0.999*
-0.895*
-1.023*
-1.690*
-0.654*
0.145
0.142
0.142
0.140
0.142
0.136
0.139
0.141
*: significant at two-sided 5% level; + significant at twosided 10% level
45
Table A.8: Estimates Model Section 6: Thresholds Equation (Eq. (8); model with DIF)
coeff.
Work
st. error
coeff.
st. error
0.097*
0.015
Work*EU
-0.104* 0.023
Female
Married/LT
Educyrs
Heart prob
Lung dis
High blood
Diabetes
Pain
Arthritis
Cancer
Cesd score
Obese
Age 58-64
Age 65-71
Age 72+
0.096*
0.051*
0.003
-0.016
-0.002
-0.024+
-0.065*
-0.038*
-0.005
0.066*
-0.022*
-0.039*
0.021
0.020
0.021
0.013
0.013
0.002
0.016
0.022
0.013
0.014
0.015
0.014
0.020
0.004
0.014
0.020
0.016
0.020
const thrh 1
thr2 - thr1
thr3 - thr2
thr4 - thr3
0
0.722*
0.704*
0.822*
0.010
0.010
0.012
sigma u
0.426*
0.007
Female*EU
Marr/LT*EU
Educyrs*EU
Heart*EU
Lung*EU
Highbl*EU
Diabetes*E
Pain*EU
Arthr*EU
Cancer*EU
Cesd*EU
Obese*EU
Age58-64*EU
Age65-71*EU
Age72+*EU
Germany
Sweden
Netherland
Spain
Italy
France
Greece
Belgium
-0.082*
-0.031
-0.004+
-0.014
0.037
0.028
0.069*
0.046*
-0.008
-0.048
0.008
0.014
0.034
0.039
0.026
-0.187*
-0.452*
-0.074+
-0.356*
-0.083*
-0.147*
-0.417*
-0.114*
0
*: significant at two-sided 5% level; + significant at twosided 10% level
46
0.019
0.020
0.002
0.027
0.033
0.020
0.025
0.022
0.024
0.036
0.008
0.023
0.028
0.027
0.030
0.047
0.051
0.043
0.049
0.042
0.043
0.046
0.046
Table A.9: Estimates Model Section 6: Vignette Equation (Eq. (5); model with DIF)
coeff.
st. error
dvigpain1
dvigpain2
dvigpain3
dvigaff1
dvigaff2
dvigaff3
dvigcvd1
dvigcvd2
dvigcvd3
1.453*
0.631*
1.598*
0.803*
0.853*
0.455*
1.075*
1.161*
1.615*
0.038
0.034
0.039
0.034
0.035
0.034
0.036
0.036
0.039
Sigmav
0.527*
0.007
*: significant at two-sided 5% level; + significant at two-sided 10%
level
47
Table A.10: Estimates Model Section 6: Employment Equation (Eq. (10))
Disability
Disability*EU
Model without DIF
coeff.
st. error
-0.516*
0.038
0.319*
0.049
Model with DIF
coeff.
st. error
-0.464*
0.036
0.272*
0.049
Constant
0.473*
0.151
0.440*
0.135
Female
-0.289*
0.059
-0.246*
0.057
Married/LT
-0.069
0.065
-0.034
0.062
Educyrs
0.030*
0.009
0.034*
0.009
Heart prob
0.015
0.076
-0.015
0.074
Lung dis
-0.102
0.109
-0.125
0.107
High blood
0.016
0.059
-0.001
0.057
Diabetes
-0.189*
0.079
-0.228*
0.076
Pain
0.131*
0.064
0.102+
0.061
Arthritis
0.041
0.062
0.031
0.060
Cancer
-0.023
0.084
0.004
0.085
Cesd score
-0.035*
0.017
-0.051*
0.016
Obese
0.223*
0.064
0.204*
0.059
Age 58-64
-0.633*
0.079
-0.620*
0.078
Age 65-71
-1.285*
0.077
-1.268*
0.073
Age 72+
-1.815*
0.101
-1.813*
0.100
Female*EU
-0.282*
0.081
-0.319*
0.079
Marr/LT*EU
-0.096
0.094
-0.125
0.092
Educyrs*EU
-0.015
0.010
-0.020*
0.009
Heart*EU
-0.021
0.128
-0.002
0.127
Lung*EU
-0.032
0.161
0.002
0.160
Highbl*EU
0.034
0.085
0.051
0.084
Diabetes*EU
0.122
0.128
0.152
0.127
Pain*EU
-0.058
0.086
-0.025
0.085
Arthr*EU
-0.066
0.101
-0.058
0.100
Cancer*EU
0.058
0.155
0.034
0.155
Cesd*EU
-0.039
0.032
-0.026
0.031
Obese*EU
-0.296*
0.097
-0.284*
0.094
Age58-64*EU
-0.399*
0.098
-0.406*
0.097
Age65-71*E
-1.288*
0.133
-1.296*
0.130
Age72+*EU
-1.407*
0.211
-1.402*
0.211
Germany
0.310
0.196
0.291
0.184
Sweden
0.852*
0.206
0.784*
0.196
Netherland
0.239
0.195
0.241
0.183
Spain
0.252
0.192
0.208
0.181
Italy
-0.129
0.197
-0.126
0.186
France
0.310+
0.189
0.300+
0.177
Greece
0.215
0.191
0.157
0.183
Belgium
0.044
0.191
0.041
0.178
Log-likelihood (of
complete model)
-99650.54
-90792.8
*: significant at two-sided 5% level; + significant at two-sided 10% level.
48
Table A.11: Estimates Model Section 6: Employment equation without objective health
conditions (Eq. (10)) & Log Likelihood
Model without DIF
coeff.
st.er.
Model with DIF
coeff.
St.er.
Health
Health*EU
-0.506*
0.282*
0.032
0.041
-0.477*
0.257*
0.031
0.040
Constant
Female
Married/LT
Educyrs
Age 58-64
Age 65-71
Age 72+
0.505*
-0.280*
-0.038
0.030*
-0.627*
-1.280*
-1.830*
0.142
0.057
0.064
0.009
0.075
0.070
0.097
0.399*
-0.239*
0.000
0.036*
-0.614*
-1.261*
-1.825*
0.141
0.057
0.063
0.009
0.075
0.070
0.096
Female*EU
Marr/LT*EU
Educyrs*EU
Age58-64*EU
Age65-71*EU
Age72+*EU
-0.313*
-0.103
-0.015+
-0.391*
-1.277*
-1.366*
0.079
0.093
0.009
0.094
0.125
0.205
-0.347*
-0.134
-0.021*
-0.398*
-1.287*
-1.365*
0.078
0.088
0.009
0.092
0.125
0.204
Germany
Sweden
Netherland
Spain
Italy
France
Greece
Belgium
0.218
0.760*
0.132
0.129
-0.255
0.175
0.084
-0.067
0.182
0.194
0.183
0.180
0.183
0.174
0.174
0.178
0.260
0.748*
0.199
0.142
-0.188
0.226
0.080
-0.006
0.181
0.193
0.182
0.179
0.181
0.172
0.165
0.177
log likelihood (of
complete model)
-99670.88
*: significant at two-sided 5% level; + significant at two-sided 10% level.
49
-90815.5
APPENDIX B. The SHARE and HRS Work Disability Vignettes and Self-Reports
SHARE
Work limiting disabilities
Three domains: affect, pain, cardio-vascular disease.
One overall self-report on work-limiting disabilities.
Three vignettes per domain.
Response categories (for all questions; same as for health domains):
1. None; 2. Mild; 3. Moderate; 4. Severe; 5. Extreme/Cannot Do
Self report:
Do you have any impairment or health problem that limits the kind or amount of work you can
do?
Introduction to vignettes:
We would now like to give you a number of examples of persons with some health problems.
We would like you to indicate the extent to which you think these people would be limited in the
kind or amount of work they can do. In terms of their age, their education, and their work
histories, you should imagine that these men or women are similar to yourself. Other than the
conditions explicitly mentioned, you should imagine the individual is in reasonably good health.
General form of vignette questions:
How much was [name of person] limited in the kind or amount of work he/she could do?
Affect Vignettes:
1. [Eva] feels worried all the time. She gets depressed once a week at work for a couple of days
in a row, thinking about what could go wrong and that her boss will disapprove of her
condition. But she is able to come out of this mood if she concentrates on something else.
2. [Tamara] has mood swings on the job. When she gets depressed, everything she does at
work is an effort for her and she no longer enjoys her usual activities at work. These mood
swings are not predictable and occur two or three times during a month.
3. [Henriette] generally enjoys her work. She gets depressed every 3 weeks for a day or two and
loses interest in what she usually enjoys but is able to carry on with her day-to-day activities
on the job.
Pain Vignettes:
1. [Yvonne] has almost constant pain in her back and this sometimes prevents her from
doing her work.
2. [Catherine] suffers from back pain that causes stiffness in her back especially at work but
is relieved with low doses of medication. She does not have any pains other than this
generalized discomfort.
50
3. [Mark] has pain in his back and legs, and the pain is present almost all the time. It gets
worse while he is working. Although medication helps, he feels uncomfortable when
moving around, holding and lifting things at work
CVD Vignettes:
1. [Norbert] has had heart problems in the past and he has been told to watch his cholesterol
level. Sometimes if he feels stressed at work he feels pain in his chest and occasionally in
his arms.
2. [Tom] has been diagnosed with high blood pressure. His blood pressure goes up quickly
if he feels under stress. Tom does not exercise much and is overweight.
3. [Dan] has undergone triple bypass heart surgery. He is a heavy smoker and still
experiences severe chest pain sometimes.
Randomization
All vignette questions are asked to all respondents. The only randomization concerns the names
of the vignette persons and the order of the vignette questions. The SHARE questionnaire had
two versions. Each has alternating male and female names on the vignettes, but where version 1
has a male name, version 2 has a female name, etc. Version 1 has the order of the questions as
given above. Version 2 reverses things: first work disability, then health; domains in reverse
order; vignettes in each domain in reverse order. The self-report always comes before the
vignettes.
HRS
Although the vignettes and self-reports in HRS are very similar to those in SHARE, there are a
few noticeable differences.
Self-report in 2004 leave-behind questionnaire:
To what extent are you limited in the kind or amount of work you can do because of an
impairment or health problem?
Not at all limited, Mildly limited, Moderately limited, Severely limited, Cannot do any work
Vignettes
HRS fielded the same vignettes as used in Kapteyn et al. (2007), with some slight changes in
formulation. The SHARE vignettes are a subset of these. Comparing the vignettes that are
common between HRS and SHARE, we find one difference. SHARE vignette CVD-2 above in
the HRS version reads:
Diane has been diagnosed with high blood pressure. Her blood pressure goes up quickly if she
feels under stress. Diane does not exercise much and is overweight. Life can sometimes be hectic
for her. She does not get along with her boss very well.
51